v0.5 — Apache-2.0

One log.
Every query shape.

kynetradb collapses search, vector, KV, document, auth, files, and realtime into a single Rust binary. One ingest. One query surface. One platform you control.

Deploy in 1 click Compare to Redis

1.07 ms BM25 @ 100k SKUs

2.21 ms Vector cosine @ 100k SKUs

18 Deploy targets

12 Outbound DB sinks

What ships in the binary

core

Universal log

Every write — entity, file, agent run, schema change — appends to a single durable WAL with fsync. Crash recovery is replay; time travel is a bookmark.

search + vector + kv + doc

Four query shapes, one process

BM25 full-text (1.07 ms @ 100k), parallel cosine vector (2.21 ms @ 100k), KV point lookup, and JSON predicate filtering. No separate services. No CDC pipelines.

auth

Multi-user auth

bcrypt-hashed passwords (cost 12) + stateless JWT sessions (HS256, 7-day TTL). Three roles: Admin / Editor / Viewer. Per-collection CRUD rules. Routes at /v1/auth/.

files

File storage

Two backends: local content-addressable layout or any S3-compatible endpoint (AWS S3, R2, MinIO, Wasabi, B2, DigitalOcean Spaces). AWS SigV4 signing built-in.

realtime

Realtime SSE

Subscribe to a live stream of writes filtered by topic and entity kind. One GET request, no client library required. Backpressure surfaces as dropped-message lag.

agents

Agentic admin

10 typed built-in actions, a real LLM tool-use loop (Anthropic + OpenAI + Ollama), persisted run audit trail, and an embedded admin SPA at /admin/. No JS build step.

Why one binary?

A typical ecommerce search stack looks like this:

Shopify  --CDC-->  Algolia (search)
   |-CDC-->  Pinecone (vector reco)
   |-CDC-->  Redis (cache)
   '-CDC-->  Postgres (analytics)

Four vendors. Four CDC pipelines. Four bills. Four failure modes. Four query languages. kynetradb collapses the stack: one ingest, one query surface, one binary, one platform you control.

When your Algolia index and your Pinecone index drift out of sync (and they will), you debug two systems simultaneously. kynetradb has one source of truth: the universal log.

Honest state (v0.5)

Real today

Single-process embeddable engine with durable WAL + crash recovery. BM25 search (case + accent folding, parallel scoring). Parallel cosine vector search. Faceted document filtering. HTTP server with auth, graceful shutdown, compression, CORS. Working Shopify connector + 9 scaffolds. Multi-user auth (bcrypt+JWT, 3 roles, log-persisted). Typed Collections with 12 field types + validation + role-scoped rules. File storage (local + SigV4 S3-compatible). Realtime SSE event bus with topic+kind filters. Hybrid LLM runtime (Ollama + Anthropic + OpenAI) with auto-discovery + fallback. 12 outbound database sinks with live async fan-out. Agentic admin: 10 typed actions, real LLM tool-use loop, persisted run audit trail, embedded HTML SPA at /admin/. Dockerfile + 1-click deploy to 18 hosting targets (10 globals + 5 Indian + 3 new globals).

Not yet

HNSW vector indexes (parallel brute force scales to ~100k vectors). tantivy-grade fuzzy + synonyms. Multi-writer clustering (single instance per WAL). Persisted secondary indexes (rebuilt from log on open). Per-tenant auth scoping. Per-provider deploy/destroy/list_instances live actions for all 10 clouds (substrate + validate are in; cloud-specific deploy APIs land in v0.7). Wiremock unit tests for db-sink write paths. GCP service-account OAuth2 validation (shape-check only today).

This is v0.5 — working code, honest README. Read the full State section before betting anything serious on it.

One log.Every query shape.